The paper proposes a Fast One-stage Unsupervised person Search (FOUS) framework for the task of unsupervised domain adaptive person search. The key highlights are:
FOUS introduces a prototype-guided labeling method to efficiently assign soft labels to unlabeled target domain samples, replacing the computationally expensive clustering algorithms used in previous methods.
FOUS designs an Attention-based Domain Alignment Module (ADAM) that can align the feature representations across domains for both detection and re-identification tasks, while also reducing the adverse impact of low-quality candidate boxes from unsupervised detection.
FOUS adopts a label-flexible training network with an adaptive selection strategy to gradually refine the coarse labels assigned by the prototype-guided method.
Without any auxiliary labels in the target domain, FOUS achieves state-of-the-art performance on two benchmark datasets, CUHK-SYSU and PRW, while significantly reducing the computational cost and inference time compared to previous methods.
Extensive experiments validate the effectiveness of the proposed components, including the attention module, prototype-guided labeling, and label-flexible training, in improving the overall performance of the unsupervised domain adaptive person search task.
Til et andet sprog
fra kildeindhold
arxiv.org
Vigtigste indsigter udtrukket fra
by Tianxiang Cu... kl. arxiv.org 05-07-2024
https://arxiv.org/pdf/2405.02832.pdfDybere Forespørgsler